Abstract: Forest fires caused by climate warming and human activities are the main disasters of subtropical forests in Guangdong in autumn and winter. We selected 116 satellite ground synchronous sampling sites in the subtropical forest area of Conghua District, Guangzhou City, Guangdong Province, to monitor and evaluate the water status of vegetation living leaves by forest fuel moisture content (FMC), and to explore the relationship between fuel water content and vegetation micro-characteristics and vegetation environmental conditions, including forest geometric characteristics (such as tree height, crown width) and forest water characteristic factors (such as NDVI, NDII). In addition to the traditional linear correlation, we also find that there is a nonlinear relationship among the characteristic variables of the forest, and the determination coefficient R² is 0.88. We have adopted three different classical machine learning algorithms, including conventional linear regression, XGboost and Gradient Boosting Regression algorithms to predict FMC. Finally, the new "comprehensive voting regression" method, which integrates the three algorithms by allocating weights, has the best correlation with the initial values of the sample points and the lowest error, so this method is used to predict the FMC of the whole region, and the prediction accuracy of the sample points can reach 86%~ 87%. The low-cost and quasi-real-time FMC index based on satellite remote sensing can provide solid theory and temporal and spatial distribution data of forest fuel moisture when formulating forest fire risk management strategies.